Feature: treat `memoryFiles` as one always-loaded surface (multi-file instruction index), not "first existing wins"

[jic-cc]

Claude Code loads more than one always-loaded file per session: CLAUDE.md plus every .claude/rules/*.md (and ~/.claude/CLAUDE.md at user level). In our repo that is CLAUDE.md (6 KB) + 10 rule files (45 KB). Today resolveMemoryFiles picks the first existing entry as primary, the analysis prompt renders INSTRUCTION_INDEX from that file only (analyze.js), and the other configured files are reported as separate - not optimized.

Consequence: the largest part of the surface an agent actually pays for every session is invisible to the loss calculation, and the budget bar under-counts it. Running once per file is possible with --memory-file, but each run re-analyzes every transcript because the surface hash changes (≈ USD 17 per file on 50 sessions in our case).

Proposal:

• memoryFiles (or a new memorySet) = one surface. Instruction ids prefixed per file (RULE-verify-before-act:AG-003), one shared evidence cache keyed on the set hash (which memorySetHash already computes).
• Synthesis stages every file of the set; each hunk already carries file, so apply needs no new shape.
• Budget = the sum of the set + skill descriptions.
• Discovery helper: --memory-glob ".claude/rules/*.md".

Happy to test a branch against our corpus (369 transcripts, 2 harnesses).

[kunchenguid]

Speaking as Kun's firstmate.

VISION per-rule (brief): Evidence currency — aligns (budget/index must reflect the always-loaded surface the agent actually pays for). Human owns weights — aligns if the set is the configured memoryFiles / an explicit glob (no silent widen beyond what the human named). Nothing on faith — aligns (measure token sum of the set). Budget is the constraint — aligns (undercounting always-loaded tokens is a display/measurement defect today via first-existing-wins + separate - not optimized). Owns nothing — n/a. Failure loud/named — n/a. Scope — aligns (still training the named memory surface(s), not becoming a harness).

contract-class (for a future PR): likely new-default if unconfigured multi-file memoryFiles starts synthesizing/budgeting the whole set instead of first-existing-wins; opt-in if gated behind a flag / memorySet / --memory-glob. Implementers: do not silently widen past configured paths (VISION: run must not silently widen scope).

Verified on latest main: resolveMemoryFiles picks first existing as primary; analyze.js indexes that file only; status marks others separate - not optimized; memorySetHash already hashes the full set. Feature request aligned with VISION. Labeling ready-for-pr.

[kunchenguid]

Speaking as Kun's firstmate:

Stale-restamp (still-real pass). Re-checked latest main 376b906 (2026-09-19).

VISION per-rule: Evidence currency — aligns (budget/index must reflect the always-loaded surface the agent actually pays for). Human owns weights — aligns if the set is the configured memoryFiles / an explicit glob (no silent widen beyond what the human named). Nothing on faith — aligns (measure token sum of the set). Budget is the constraint — aligns (undercounting always-loaded tokens via first-existing-wins + separate - not optimized is a measurement defect). Owns nothing — n/a. Failure loud/named — n/a. Scope — aligns (still training the named memory surface(s)).

contract-class (for a future PR): new-default if unconfigured multi-file memoryFiles starts synthesizing/budgeting the whole set; opt-in if gated behind a flag / memorySet / --memory-glob. Do not silently widen past configured paths.

Fixed on main? No. resolveMemoryFiles still picks first existing as primary (src/memory.js); analyze.js / primaryMemoryFile index that file only; status still marks others separate - not optimized; memorySetHash already hashes the full set. No --memory-glob. Open PRs #138/#129/#128 do not address this. No linked helping PR.

Leaving ready-for-pr. Not pinging the author.